CheXT: Knowledge-Guided Cross-Attention Transformer for Abnormality Classification and Localization in Chest X-rays

Classical chest X-ray analysis has designed radiomic features to indicate the characteristics of abnormality of the chest X-rays. However, extracting reliable radiomic features heavily hinges on pathology localization, which is often absent in real-world image data. Although the past decade has witnessed the promising performance of convolutional neural networks (CNNs) in analyzing chest X-rays, most of them ignored domain knowledge such as radiomics. Recently, the surge of Transformers in computer vision has suggested a promising substitute for CNNs. It can encode highly expressive and generalizable representations and avoid costly manual annotations via a unique implementation of the self-attention mechanism. Moreover, Transformers naturally suit the feature extraction and fusion from different input modalities. Inspired by its recent success, this paper proposes \textbf{CheXT}, the first Transformer-based chest X-ray model. CheXT targets (semi-supervised) abnormality classification and localization from chest X-rays, enhanced by baked-in auxiliary knowledge guidance using radiomics. Specifically, CheXT consists of an image branch and a radiomics branch, interacted by cross-attention layers. During training, the image branch leverages its learned attention to estimate pathology localization, which is then utilized to extract radiomic features from images in the radiomics branch. Therefore, the two branches in CheXT are deeply fused and constitute an end-to-end optimization loop that can bootstrap accurate pathology localization from image data without any bounding box used for training. Extensive experiments on the NIH chest X-ray dataset demonstrate that CheXT significantly outperforms existing baselines in disease classification (by 1.1\% in average AUCs) and localization (by a \textbf{significant average margin of 3.6\%} over different IoU thresholds). Codes and models will be publicly released.